Using deep learning to deconvolute complex spectra for hyperspectral imaging applications

August 05, 2019

Samantha Rudinsky (1), Yu Yuan (1), Francis B. Lavoie (2), Raynald Gauvin (1), Ryan Gosselin (2), Nadi Braidy (2), Nicolas Piché (3), Mike Marsh (4)
Microscopy and Microanalysis, 25, Supplement 2, August 2019: 178-179. DOI: 10.1017/S1431927619001624


X-ray energy dispersive spectroscopy (EDS) for materials characterization is one of most widespread analytical methods. When EDS is applied in rastering spatial patterns, the aggregate spectra yield hyperspectral images (HSI) that provide a quick and precise interpretation of the elemental spatial distribution of nanomaterials. However the long dwell time required for acquisition of each spectrum makes the experiments highly resource-intensive. We discuss here a Deep Learning approach that uses neural networks to deconvolute overlapping spectra having a weak signal-to-noise ratio, thereby making more precise and definite interpretation possible with shorter imaging times and electron doses.

Author Affiliation

(1) Department of Materials Engineering, McGill University. Montreal, Canada.
(2) Department of Chemical Engineering and Biotechnological Engineering, Université de Sherbrooke. Sherbrooke, Canada.
(3) Object Research Systems. Montreal, Canada.
(4) Object Research Systems. Denver, USA.